DeepSeek V4.1 Flash 1017 is out: what changed, pricing and API access
DeepSeek V4.1 Flash 1017 is now on Dreamtune. Explore the October 17 update, official benchmark results, current model pricing and how to try it.
Now available in Dreamtune: DeepSeek V4.1 Flash 1017 is listed in the live model catalog under deepseek/deepseek-v4.1-flash-1017.
DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, updating its Flash model across the official API, website and app. The API model name remains deepseek-flash, and DeepSeek says this update leaves its API pricing unchanged. This is an update to V4.1 Flash, which launched in September. Source: DeepSeek’s official changelog.
For developers, the useful question is whether it completes your tasks more accurately, with fewer tokens and less waiting. Here is what the release establishes and how we would evaluate it. Sources and prices were checked on October 18, 2026. We have not independently benchmarked this release.
What changed in DeepSeek V4.1 Flash 1017?
DeepSeek reports improvements across reasoning, coding, agent tasks and visual understanding, alongside shorter task runtimes and outputs. Those are the vendor’s findings; they do not establish the same gains for every application. The release also includes optimizations for particular agent harnesses. Source: October 17 release notes.
The model card gives several comparisons against the preceding V4.1 Flash:
- Terminal-Bench 4.0: 42.3, up from 31.2.
- DeepSWE v1.1: 77.0, up from 74.2.
- ProgramBench: 29.1, up from 20.3.
- Automation-Bench v1.0.6: 60.5, up from 54.8.
These are DeepSeek-reported benchmark scores. The card identifies Claude Code as the Terminal-Bench harness and Mini-SWE-V2 as the DeepSWE harness. Results depend on that setup, so compare them within the same benchmark rather than treating a score as a universal success rate. The model weights are published under the MIT license. Source: official model card.
Our interpretation: the release is a useful candidate for another evaluation of coding and tool-heavy workflows. It is not evidence that an existing application should switch without testing.
What does the DeepSeek API cost?
DeepSeek’s direct API lists the following Flash prices per one million tokens, checked October 18:
- Cached input: $0.006 at peak times; $0.003 off-peak.
- Uncached input: $0.30 at peak times; $0.15 off-peak.
- Output: $1.20 at peak times; $0.60 off-peak.
DeepSeek’s pricing page defines the peak and off-peak schedule. These are direct-provider rates, not a Dreamtune price quote. Recheck them before budgeting. Source: DeepSeek Models & Pricing.
For example, 100,000 uncached input tokens and 20,000 output tokens cost $0.054 at peak rates or $0.027 off-peak, using that table. This calculation excludes cache hits and separately billed services. An agent’s repeated calls all contribute to its total.
Which model name should developers use?
On DeepSeek’s own API, use deepseek-flash. Its pricing page maps that identifier to DeepSeek-V4.1-Flash-1017 and lists a one-million-token context window, vision, tool calls and thinking support. Other gateways can have different identifiers, release timing and feature support. Source: current model details.
When evaluating, record both the identifier you requested and the version your provider says it serves. An unchanged alias can point to an updated model. Keep the evaluation date with your results so a future comparison remains meaningful.
Thinking mode is enabled by default on DeepSeek’s direct API, with high effort. If response time or cost changes unexpectedly, check your selected thinking settings before attributing everything to the model update. Source: thinking-mode documentation.
What should you test before switching an agent?
Use a small collection of real tasks with outcomes you can check. Our suggested starting set is a code change, a factual research question, a structured extraction and a task that requires a tool.
For each, record:
- Whether the final result meets the task’s acceptance criteria.
- Whether tool arguments are correct and the resulting evidence supports the answer.
- Total elapsed time and input/output tokens across the whole task.
- Total cost, including failed attempts and additional services.
- Whether a person had to intervene or repair the output.
Keep the task, tool access and effort settings comparable. A shorter answer helps only if it still contains what the user needs. A faster first response helps less if completing the task takes several extra turns.
For a repeatable comparison process, see how to choose an AI model and how to make your AI agent better.
Can I use the new DeepSeek model through Dreamtune?
Yes. Dreamtune’s live model catalog lists deepseek/deepseek-v4.1-flash-1017 for text conversations. The listing was checked on October 18, 2026. Choose that exact model to use the October 17 release.
At that check, the catalog listed $0.15 per million input tokens and $0.60 per million output tokens. Dreamtune’s listed rates can differ from DeepSeek’s direct API schedule above; check the catalog before running a request. Availability of an upstream feature does not automatically establish support for that feature in every Dreamtune interface.
Dreamtune provides access to supported models, tools and GPU compute through one account and prepaid balance. Start with the connection guide to see how an existing agent can use those services. Check pricing for current Dreamtune fees and offers.